Posture monitoring method, posture monitoring device, posture analysis method and posture analysis system

The present application claims priority to Singapore App. No. SG10202400954V filed on Apr. 1, 2024 with the Intellectual Property Office of Singapore, the entire disclosure of which is incorporated herein by reference.

The present disclosure relates to posture monitoring and analysis. In particular, the present disclosure relates to the wearable devices and the use of machine learning to monitor and analyze posture.

Workplace-related musculoskeletal disorders, particularly back injuries, are among the most common and costly occupational health issues. Industries such as logistics, warehousing, and healthcare rely on manual handling tasks, increasing the risk of chronic injuries due to repetitive stress and improper posture.

Current ergonomic solutions, including motion-tracking wearables and manual assessments, suffer from limitations such as lack of real-time feedback to correct posture dynamically, an inability to quantify and track high-risk movements with precision and limited adaptability to different body types and job roles.

generating an alert based on the classification of the posture of the wearer. According to a first aspect of the present disclosure, a posture monitoring method is provided. The method comprises: measuring force sensor data from at least four force sensors positioned on a wearable device; determining pressure distribution data from the force sensor data, the pressure distribution data indicating a distribution of pressure across a lumbar region of a wearer of the wearable device; using a machine learning model system to classify a posture; and

The method analyzes movement patterns and detects improper posture in real time. Posture classifications may include Good pickup, Upright, Forward Bend, Backward Bend, Right Twist and Left Twist. The alert may be generated if the posture of the wearer corresponds to a high-risk posture. The alert may continue until the wearer moves to a low-risk posture. The alert may be provided as an audio alert, a visual alert such as a flashing light and/or a vibration alert.

In an embodiment, the machine learning model system comprises a pressure data to posture model trained to classify the posture the wearer from the pressure distribution data.

In an embodiment, the machine learning model system comprises a pressure data to pose model trained to estimate a pose of the wearer of the wearable device and a pose to posture model trained to classify a posture of the wearer based on the estimated pose, and wherein using the machine learning model system to classify a posture of the wearer comprises estimating the pose of the wearer using the pressure data to pose model and classifying the posture of the wearer using the pose to posture model.

The pressure data to pose model may be configured to generate an intermediate representation that encodes information related to the pose of the wearer. The intermediate representation may take various forms, including but not limited to: a set of estimated key skeletal points, a vector representation of the pose, joint angle configurations, kinematic chains, or latent embeddings produced by a neural network. The specific form of the intermediate representation may be selected based on the requirements of the downstream application or processing pipeline.

The pressure distribution data may be determined from the force sensor data using baseline calibration data which may be determined during a calibration period in which the wearer of the wearable device adopts a calibration posture.

The method may comprise, during a device fit calibration period, comparing an average sensor output from the at least four force sensors with a calibration window and generating an indication indicating a result of the comparison.

According to a second aspect of the present disclosure, a wearable device for monitoring posture is provided. The wearable device comprises: a sensor array comprising at least four force sensors arranged to be positioned on a lumbar region of a wearer of the wearable device; a data storage unit storing a machine learning model system configured to classify a posture of the wearer of the wearable device using pressure distribution data indicating a distribution of pressure across the lumbar region of the wearer; a feedback unit configured to generate alerts; and a controller configured to receive force sensor data from the at least four force sensors, determine pressure distribution data from the force sensor data, input the pressure distribution data into the machine learning model system, and to control the feedback unit to generate an alert if the posture of the wearer is classified based on the classification of the posture of the wearer.

In an embodiment, the machine learning model system comprises a pressure data to posture model trained to classify the posture the wearer from the pressure distribution data.

In an embodiment, the machine learning model system comprises a pressure data to key points model trained to estimate a pose of the wearer of the wearable device and a pose to posture model trained to classify a posture of the wearer based on the estimated pose, and wherein using the machine learning model system to classify a posture of the wearer comprises estimating the pose of the wearer using the pressure data to pose model and classifying the posture of the wearer using the pose to posture model.

The wearable device may comprise a belt portion configured to be worn around the waist and being configured to adjust a fit of the wearable device on the wearer, wherein the controller is further configured to compare an average sensor output from the at least four force sensors with a calibration window during a device fit calibration period and to control the feedback unit to generate an indication of a result of the comparison.

The controller may be further configured to determine pressure distribution data from the force sensor data by determining a deviation of the force sensor data from baseline calibration data. In an embodiment, the controller is further configured to determine the baseline calibration data during a sensor calibration period in which the wearer of the wearable device adopts a calibration posture.

The wearable device may further comprise a tilt sensor configured to generate orientation data, wherein the controller is further configured to input the orientation data into the machine learning model system.

The wearable device may further comprise a wireless communication module configured to generate signals indicative of the pressure distribution data and corresponding time stamp data. The signals may be sent to a server for posture analysis of the wearer over a period of time.

According to a third aspect of the present disclosure, a posture analysis method is provided. The posture analysis method comprises: receiving pressure distribution data and corresponding time stamp data, the pressure distribution data indicating a distribution of pressure across a lumbar region of a wearer of a wearable device; using a machine learning model system to classify a posture of the wearer into one of a set of posture types at each of a plurality of times indicated by the time stamp data; and generating a posture analysis report indicating a time spent by the wearer in posture types of the set of posture types.

In an embodiment, the machine learning model system comprises a pressure data to posture model trained to classify the posture the wearer from the pressure distribution data.

In an embodiment, wherein the machine learning model system comprises a pressure data to pose model trained to estimate a pose of the wearer of the wearable device and a pose to posture model trained to classify a posture of the wearer based on the estimated pose, and wherein using the machine learning model system to classify a posture of the wearer comprises estimating the pose of the wearer using the pressure data to pose model and classifying the posture of the wearer using the pose to posture model.

According to a third aspect of the present disclosure, a posture analysis system is provided. The posture analysis system comprises: a processor, a machine learning model storage device storing a machine learning model system and a data storage device storing computer program instructions operable to cause the processor to: receive pressure distribution data and corresponding time stamp data, the pressure distribution data indicating a distribution of pressure across a lumbar region of a wearer of a wearable device; use a machine learning model system to classify a posture of the wearer into one of a set of posture types at each of a plurality of times indicated by the time stamp data; and generate a posture analysis report indicating a time spent by the wearer in posture types of the set of posture types.

In an embodiment, the machine learning model system comprises a pressure data to posture model trained to classify the posture the wearer from the pressure distribution data.

In an embodiment, the machine learning model system comprises a pressure data to pose model trained to estimate a pose of the wearer of the wearable device and a pose to posture model trained to classify a posture of the wearer based on the estimated pose, and wherein using the machine learning model system to classify a posture of the wearer comprises estimating the pose of the wearer using the pressure data to pose model and classifying the posture of the wearer using the pose to posture model.

The present disclosure relates to posture monitoring and analysis. As described in more detail below, a wearable device is provided with sensors which allow real-time monitoring of the posture of a wearer of the device and the generation of alerts when the wearer enters a high risk posture. In addition to real-time monitoring, the posture of the wearer may also be analyzed over periods of time.

1 FIG. 1 FIG. 100 110 200 100 200 120 120 200 100 is a schematic of a system comprising a wearable device and a posture analysis system according to an embodiment of the present invention. As shown in, a wearable deviceis provided with a belt portionand can be worn by a wearer. When the wearable deviceis worn by the wearer, force sensorsarranged on the wearable devicemeasure a pressure distribution on the lower back or lumbar region of the wearer. Using the pressure distribution, a machine learning model provided in the wearable deviceclassifies the posture of the wearer. The classification may be to classify the posture as either high-risk or low risk, or may be to classify the posture from among a set of posture types such as good pickup, upright, forward bend, backward bend, right twist and left twist. If the posture is classified as high-risk an alert may be generated to alert the wearer of the high-risk posture. The alert may be, for example, an audio alert, a visual alert or a vibration alert.

100 The wearable deviceuses multi-point force sensors to capture real-time pressure, strain, and tilt from the user's back. These sensors are embedded in a flexible yet durable material that conforms to various body types while maintaining high sensitivity to force distribution.

100 210 220 700 700 200 230 230 240 200 In addition to providing real-time monitoring and alerts, the wearable devicemay also send pressure distribution datato a communication devicesuch as a smart phone device which in turn sends the pressure distribution data to a posture analysis system. The posture analysis systemmay be implemented as a server or a cloud based system which analyses the posture of the wearerover time and generates posture analysis reports. The posture analysis reportsmay be sent to a user devicewhich may, for viewing, by, for example a supervisor of the wearer.

2 FIG.A 2 FIG.D 2 FIG.A 2 FIG.A 130 6 132 132 130 toshow a wearable device according to an embodiment of the present invention.shows the front side of the wearable device which faces the lumbar region of the wearer when the device is worn. As shown in, the front side of the wearable device has a body portionon whichsensor padsare formed. The sensor padsare formed from a gel material having a Shore hardness of 0-10. The thickness of the gel pads is 10 mm or more. The body portionis formed from a soft elastomer to allow the wearable device to accommodate different body shapes.

2 FIG.B 2 FIG.B 134 136 134 138 134 shows the rear side of the wearable device which faces away from the wearer when the device is worn. As shown in, the rear side of the wearable device has a body portionon which connectorsfor the belt portion are arranged. The body portionalso comprises a housingwhich houses control circuitry of the wearable device. The body portionis formed from a plastic material which provides rigidity to ensure accurate pressure readings.

2 FIG.C 2 FIG.C 140 132 134 140 140 6 132 shows an exploded view of the wearable device. As shown in, a sensor circuitis provided between the body portionof the front surface and the body portionof the rear surface of the wearable device. The sensor circuitis provided on an IP68 waterproof breathable membrane. The sensor circuitcomprisesforce sensors which are located under the gel padson the front surface of the wearable device.

2 FIG.D 2 FIG.D 5 FIG. 110 140 142 shows the adjustment of the belt portion of the wearable device. As shown in, the belt portionof the wearable device is provided with an adjustable fastener. The adjustable fastener may be a double ringor friction buckle. The adjustment allows the pressure on the sensor array to be calibrated. The calibration process is described in more detail below with reference to.

3 FIG. 3 FIG. 1 FIG. 150 152 154 156 158 160 152 154 156 700 160 162 162 162 158 156 154 158 120 148 120 6 is a block diagram showing control circuitry of a wearable device according to an embodiment the present invention. As shown in, the wearable device control circuitrycomprises a power module, a feedback unit, a wireless communication module, a controllerand a data storage unit. The power modulecomprises a battery which supplies power to the wearable device. The feedback unitmay comprise an actuator to provide vibration feedback, a speaker or buzzer to provide audio feedback, a light or group of lights to provide visual feedback or a combination feedback types. The wireless communication modulemay be Bluetooth or WiFi module which allows the wireless device to send and receive data wirelessly, for example to send pressure distribution data to the posture analysis systemshown in. The data storage unitstores a machine learning system. The machine learning systemis configured to classify the posture of the wearer of the wearable device based on the pressure distribution data. In some embodiments, accelerometer data or orientation data is also input into the machine learning model systemto classify the posture of the wearer of the wearable device. The controlleris a microcontroller or system on a chip (SoC) which performs processing and controls the wireless communication moduleand feedback unit. The controlleris also connected to the pressure sensor arrayand an accelerometer. The pressure sensor arraymay provided as an array of piezoelectric force sensors. In some embodiment, the pressure sensor array comprisessensors in a 3×2 array with two vertical rows of three sensors arranged on the lumbar region of the wearer. In other embodiments, fewer or more sensors may be provided. In order to provide an array map of back pressure distribution on the wearer, at least four sensors are required, for example in a 2×2 array. Increasing the number of sensors in the array may increase the accuracy of posture classification.

4 FIG. 4 FIG. 3 FIG. 4 FIG. 400 150 is a flowchart showing a posture monitoring method according to an embodiment of the present invention. The methodshown inis carried out by the wearable device control circuitryshown in. The method shown inmay be repeated continuously, for example at a frequency of 100 Hz.

402 120 158 In step, the pressure sensor arraymeasures force sensor data which is indicative of the forces on each of the force sensors. The force sensor data is received by the controller.

404 158 In step, the controllerdetermines pressure distribution data from the force sensor data. The pressure distribution data may be determined from the force sensor data by determining a deviation from baseline calibration force data and multiplying by a predetermined constant. The calibration period may take place after the wearer puts on the wearable device and the wearer may be directed to adopt a calibration posture for a calibration period.

406 162 158 162 In step, the pressure distribution data is input into the machine learning model systemby the controller. The machine learning model systemclassifies the posture of the wearer. The classification of the posture of the wearer may be into one of a set of classification or may be a classification as high-risk or low-risk.

408 158 154 In step, the controllercontrols the feedback unitto generate an alert based on the classification of the posture of the wearer. For example, the alert may be generated if the posture of the wearer is classified as high risk. The alert may be generated as haptic, sound and/or light.

A de-calibration process may be implemented as follows. If, after calibration, the average sensor values drop below a defined threshold and remain below that threshold for a specified period, the device automatically de-calibrates. Upon de-calibration, the device produces a brief vibration and plays a calibration tone to alert the user that the calibration state has been reset, prompting a re-calibration if necessary.

5 FIG. 5 FIG. 3 FIG. 500 150 500 is a flowchart showing a method of calibrating the fit of a wearable device according to an embodiment of the present invention. The methodshown inis carried out by the wearable device control circuitryshown in. The methodis carried out when the wearer puts on the wearable device and is carried out to calibrate the fit of the wearable device. For example, after the user puts on the device, they may be instructed to adjust the device to a predefined tightness while adopting a specific calibration posture.

502 120 158 The wearer puts on the wearable device and following the wearer putting on the wearable device, in step, the pressure sensor arraymeasures force sensor data which is indicative of the forces on each of the force sensors. The force sensor data is received by the controller.

504 158 500 506 508 510 512 514 516 5 FIG. In step, the controllercalculates an average sensor reading and compares the average sensor reading with a calibration window. The calibration window indicates a range of force sensor readings. The methodmoves to stepif the average sensor reading is below the calibration window. This would occur if the belt of the wearable device is too loose. The method then moves to stepand in this example no feedback is given which indicates that the fit is too loose. If the average sensor reading is within the calibration window, the method moves to step, and then to stepin which intermittent feedback is given to the wearer to indicate that the fit is correct. If the average sensor reading is above the calibration window, the method moves to step, and then to stepin which stronger or faster feedback is given to the wearer to indicate that the fit is too tight. It will be appreciated that the feedback shown inis one example and other types of feedback may be given for the three results of the comparison of the average sensor reading with the calibration window.

Once the average sensor value remains within the calibration window for a set period (e.g., 3 seconds), the device: provides confirmation feedback via a brief vibration and a calibration tone and captures and stores the sensor values at that moment as baseline calibration data (baseline force readings) for subsequent posture monitoring.

2 FIG.D As described above, feedback is given to the wearer to indicate whether the fit is in the correct range and if the fit is outside the correct range, the wearer can adjust the fit using the adjustable fastener shown in.

Prior to calibration, the controller may continuously sample force sensor data and maintains a sliding window of these values. Then during the calibration/tightness determination phase, the average value of the sensor data within this window is computed in real time to evaluate the device's tightness. When the device remains within the calibration range for a set duration (e.g., 3 seconds), the controller captures the average sensor values from the final state of the window as the calibration baseline. Once calibrated, the controller increases the sampling rate (e.g., to 100 Hz) for enhanced precision in monitoring. The controller continuously calculates the average of the real-time sensor data and compares it against the stored calibration baseline multiplied by a predetermined constant (derived from extensive testing). This comparison determines different postures or pressure distribution states, thereby triggering the appropriate feedback

6 FIG. 3 FIG. 7 FIG. 600 162 732 is a block diagram showing a machine learning model system in a wearable device or a posture analysis system according to an embodiment of the present invention. The machine learning systemmay correspond to the machine learning model systemshown inor the machine learning model systemshown in.

6 FIG. 600 602 602 As shown in, the machine learning model systemtakes pressure distribution dataas input. The pressure distribution dataindicates the In some embodiments, the input also comprises accelerometer data.

602 The pressure distribution datais input into three machine learning models.

610 610 612 A pressure distribution to key points modelis trained to convert pressure distribution data into estimated skeletal key points of the wearer. The output from the pressure distribution to key points modelis fed into a key points to posture modelwhich classifies posture based on the extracted skeletal key points.

622 622 620 620 620 A posture prediction modelis trained on both key point data and pressure distribution data (and may also be trained on accelerometer data). The input to the posture prediction modelmay be fed into an attention mechanismwhich refines predictions by prioritizing the most relevant input data, for example by prioritizing particularly relevant regions in the pressure sensor data. The attention mechanismenhances model accuracy by integrating both key points and tactile data features. The attention mechanismuses an attention-based AI mechanism to weigh significant pressure distribution and key point features.

630 A pressure distribution to posture modelis trained to directly map pressure distribution data onto posture classification.

612 622 630 The outputs from the three models (the key points to posture model, the posture estimation modeland the pressure distribution model) are fed into a voting or fusion mechanism which uses ensemble techniques, including voting or fusion-based decision-making, to finalize posture classification. This enhances reliability by aggregating multiple machine learning model outputs.

600 642 The final output of the machine learning model systemis a posture classification. This classification may be a classification into one of a set of pre-defined categories (for example forward bend, backward bend, twisting, proper lift, upright, etc.). Alternatively, the classification output may be a classification as a high-risk posture or a non-high risk posture. The output posture classificationsupports both real-time and post-analysis applications for ergonomic assessment.

630 The pressure distribution to posture modelbypasses key point extraction and directly predicts posture from pressure sensor data. It is trained using: Paired datasets of pressure sensor readings and corresponding labeled postures. The wearable device collects data either wirelessly in real-time or from stored sensor data, enabling a robust training process through supervised learning. The model is trained on real-world movement data, focusing on key postures such as Upright, Forward Bend, Backward Bend, Right Twist & Left Twist), ensuring it learns posture patterns purely from pressure sensor data inputs without relying on external motion capture or visual data. This pathway allows fast and efficient posture classification without requiring intermediate skeletal mapping.

610 610 The pressure distribution to key points modelprovides for a more structured understanding of movement. The modelis trained using: pressure sensor data aligned with motion capture outputs. Human body key points extracted from video analytics (e.g., OpenPose, OpenMM). By learning the correlation between pressure distribution and skeletal movement, the model estimates key points purely from pressure sensor data input.

612 The key points to posture modelclassifies postures based on estimated skeletal key points. Training includes: Video-based movement recordings paired with key point extractions and manual labeling of posture categories. This approach allows a structured, pose-driven classification of posture, independent of direct pressure sensor inputs.

622 620 The posture estimation modeltakes both key point data and pressure distribution data as inputs. To refine predictions, the attention mechanismis employed to Identify key pressure regions that contribute most to posture classification, reduces noise from less informative sensor inputs and enhance accuracy by dynamically adjusting focus on high-impact pressure points.

6 FIG. Whileshows three processing pipelines to obtain the final posture classification, in some embodiments a smaller number of processing pipelines may be used. For example, embodiments may be implemented with only the pressure distribution to posture model. Such a configuration would reduce the processing requirements, but may have less accuracy and generalization ability. It has been found that using a two-step model (pressure data to key points followed by key points to posture classification) enables better generalization across different body types, movements, and sensor variations, making the system more robust and scalable. The two-step model may be implemented as a pressure data to pose model which generates an intermediate representation that encodes information related to the pose of the wearer. The intermediate representation may take various forms, including but not limited to: a set of estimated key skeletal points, a vector representation of the pose, joint angle configurations, kinematic chains, or latent embeddings produced by a neural network. The specific form of the intermediate representation may be selected based on the requirements of the downstream application or processing pipeline.

7 FIG. 700 710 712 714 716 720 730 110 120 712 710 730 700 730 700 700 714 700 100 716 700 716 700 716 is a block diagram showing a posture analysis system according to an embodiment of the present invention. The posture analysis systemcomprises a processor, a working memory, a network interface, an input/output interface, program storage, and data storage. The processormay be implemented as one or more central processing unit (CPU) chips. The program storageis a non-volatile storage device such as a hard disk drive which stores computer program modules. The computer program modules are loaded into the working memoryfor execution by the processor. The data storageis a non-volatile storage device which stores data which is used by the posture analysis systemduring processing. The data storagemay be implemented as a number of separate data storage devices. These devices may physically form part of the posture analysis systemor may be separate systems such as servers which communicate with the posture analysis systemto provide data when required. The network interfaceis an interface that allows the posture analysis systemto communicate with other devices and receive signals from the wearable device. The input/output interfaceallows a user to input data or commands which are processed by the posture analysis system. The input/output interfacealso allows the results of processing by the posture analysis systemto be output to the user. The input/output interfacemay be implemented as an input device such as a keyboard, pointing device or touch screen and an output device such as a display or a printer.

720 722 724 710 720 700 7 FIG. The program storagestores a posture classification module, and a posture analysis module. The computer program modules cause the processorto execute various processing methods which are described in more detail below. The program storagemay be referred to in some contexts as computer readable storage media and/or non-transitory computer readable media. As depicted in, the computer program modules are distinct modules which perform respective functions implemented by the posture analysis system. It will be appreciated that the boundaries between these modules are exemplary only, and that alternative embodiments may merge modules or impose an alternative decomposition of functionality of modules. For example, the modules discussed herein may be decomposed into sub-modules to be executed as multiple computer processes, and, optionally, on multiple computers. Moreover, alternative embodiments may combine multiple instances of a particular module or sub-module. It will also be appreciated that, while a software implementation of the computer program modules is described herein, these may alternatively be implemented as one or more hardware modules (such as field-programmable gate array(s) or application-specific integrated circuit(s)) comprising circuitry which implements equivalent functionality to that implemented in software.

700 700 700 700 Although the posture analysis systemis described with reference to a computer, it should be appreciated that the posture analysis systemmay be formed by two or more computers in communication with each other that collaborate to perform a task. For example, but not by way of limitation, an application may be partitioned in such a way as to permit concurrent and/or parallel processing of the instructions of the application. Alternatively, the data processed by the application may be partitioned in such a way as to permit concurrent and/or parallel processing of different portions of a data set by the two or more computers. In an embodiment, virtualization software may be employed by the posture analysis systemto provide the functionality of a number of servers that is not directly bound to the number of computers in the posture analysis system. In an embodiment, the functionality disclosed above may be provided by executing the application and/or applications in a cloud computing environment. Cloud computing may comprise providing computing services via a network connection using dynamically scalable computing resources. A cloud computing environment may be established by an enterprise and/or may be hired on an as-needed basis from a third party provider.

730 732 732 600 6 FIG. The data storagestores a machine learning model system. The machine learning model systemmay correspond to the machine learning model systemdescribed above with reference to.

8 FIG. 8 FIG. 7 FIG. 800 700 is a flow chart showing a posture analysis method according to an embodiment of the present invention. The methodshown inis carried out by the posture analysis systemshown in.

802 700 700 1 FIG. 1 FIG. In step, the posture analysis systemreceives pressure distribution data and time stamp data. The pressure distribution data and time stamp data may be received from a wearable device such as that shown in. As shown in, the data may be transmitted from the wearable device to a communication device before being transmitted from the communication device to the posture analysis system.

804 710 700 722 732 In step, the processorof the posture analysis systemexecutes the posture classification moduleto classify the posture of the wearer of the wearable device. As discussed above, the machine learning model systemis trained to classify the posture of the wearer based on pressure distribution data as input.

806 710 700 724 In step, the processorof the posture analysis systemexecutes the posture analysis moduleto generate a posture analysis report. The posture analysis report may indicate the postures of the wearer and the time the wearer has spent in each posture.

As described above, the present disclosure provides a wearable device for monitoring of posture of a wearer. The functionality of the device may be integrated into safety vests, belts, or back braces. The Multi-point force sensors capture real-time pressure and strain. Flexible & rigid sensing adapts to various body types. Optimized materials (0-10 Shore A gel hardness, 10 mm+ thickness) enhance sensing accuracy. The wearable device may have breathable design with perforations enables IP68 rating for durability and comfort. The calibration stage ensures accurate adjustments for various body sizes and shapes.

Whilst the foregoing description has described exemplary embodiments, it will be understood by those skilled in the art that many variations of the embodiments can be made within the scope and spirit of the present invention.